Don't Let AI Meeting Notes into the Knowledge Base

I don’t think AI-generated meeting notes should go straight into a knowledge base without being checked, especially when other AI systems will later use those documents to answer questions, write proposals, and carry out tasks.

The problem is that something that was only discussed in a meeting can end up written down as something that was decided.

Say a meeting includes this exchange:

“Should we launch next week?”

“Technically we could, but the tests haven’t finished running.”

“Then let’s see the test results and decide on Friday.”

“Sounds good.”

Once AI turns this into notes, it might come out as:

The meeting decided to launch next week.

That final “Sounds good” agreed to waiting until Friday to decide. But drop the condition before it, and it can be read as agreeing to launch next week.

It feels to me a bit like flattening something three-dimensional into two dimensions.

In a meeting, every remark is attached to who said it, whom they were responding to, their tone, the conditions they set, and who actually has the authority to decide. In the notes, all that may be left is a few neatly lined-up conclusions.

When the tech lead says “we could,” they may only be confirming that it’s technically doable. When the person who owns the schedule says “let’s hold off,” it means the timing isn’t settled yet. Who is proposing, who is judging feasibility, and who makes the final call all affect how a remark should be understood.

Meetings also move fast, with several people picking up on one another’s points, and AI may not even keep straight who said what. Even if it does, it won’t necessarily know how those people divide the work, or who has the authority to decide this particular matter. The recording doesn’t always spell out that kind of background.

So even if every sentence is transcribed correctly, “we talked it over” can still turn into “the team reached consensus” once it’s summarized. Nobody misheard the words; the notes still got the facts wrong.

People who were in the meeting can read the notes and fill in from memory: “That was only brought up. Nothing was decided.”

A model reading the knowledge base later, if all it gets is the notes, has no such memory. What it sees is a document with a formal title, a clear structure, and a confident tone, stating that “the meeting decided to launch next week.”

If you tell it to work strictly from the knowledge base, it may diligently take those notes as its basis: arrange the release process, prepare the customer notice, and put the launch date into the project plan.

It doesn’t even need to make anything up along the way. Simply by trusting those notes, it can carry an unconfirmed proposal into a whole chain of follow-up work.

This is also why I think trustworthy data will matter more as models get stronger. A model can make excellent use of its premises and do the follow-up work thoroughly, but whether those premises were ever actually confirmed is a separate question.

Of course, a stronger model might also notice contradictions and go back to verify. But if “decide on Friday” has already been cut, and the original recording, the speakers’ identities, and any record of confirmation are all gone, the notes alone may not tell it that the matter is still open.

The data quality I mean here also covers words that look safe to cut: “maybe,” “tentative,” “to be confirmed,” “if the tests pass.” These words determine whether a piece of information can serve as a basis for later work. Remove them to make the notes more concise, and you may change the status of the thing itself.

When I say “don’t let them into the knowledge base,” I mean unconfirmed AI meeting notes shouldn’t go straight to being company facts that later AI systems trust by default. The notes can be kept for reference, but proposals, items awaiting confirmation, and formal decisions should each keep their own status. When a confirmed conclusion goes into a formal document, it should carry its original conditions with it.

Back to the example: what the meeting actually arranged was to look at the test results first and decide on Friday. As for launching next week, someone suggested it, but it hasn’t been decided.